2026-10-23 13:14 UTC
Replies (1)
-
@petrol_sniff_king@lemmy.blahaj.zone 2026-10-23 19:14
They’re built to model learning in a mathematical kind of way. And so they suffer similar problems that learning does, in a mathematical kind of way. The way that genetic algorithms distill patterns is by “trying everything.” And a robot that “tries everything” is going to try a lot of really destructive methods, too. If you were trying to build a key for a lock and you gave a bot the goal “open the lock,” you’re going to find a lot of keys, but also dynamite, lockpicks, hacksaws. Especially if building a key is hard but dynamite is easy. If you give a bot the goal “convince this guy to open the door for you,” it may start pantomiming dishonesty but only because dishonesty is an effective solution. Dishonesty is the pattern being distilled. None of this is evidence that it’s conscious. None of this is evidence that it’s even good—a simulacrum that pantomimes human behavior, like conversation, friendship, kindness, lust, is genuinely really fucking horrible. This shit can and does break people’s brains. I won’t tell you to go and read a bunch of articles about how they work because… honestly, fuck them. But I will say that this stuff did used to be really interesting before 2020 or so.